=======================
Learning with Curricula
=======================
- Joseph O'Sullivan
I have been interested in improving learning accuracy when multiple
functions are learned over time and particularly interested in
exploiting some notion of curriculum in such scenarios. Three coupled
questions arise: (1) what algorithms can be developed that benefit from
previous learning experience when presented with a new function to be
learned? (2) what curriculum, or ordering of functions, produces the
most effective learning? and (3) how should such curricula be used?
In this talk, I'll define "Learning with Curricula" and introduce a
notation for discussing it. I'll present two specific algorithms that
learn with curricula, SMTL and STINT. SMTL is an extension of MTL that
allow us to transfer knowledge sequentially from multiple previously
learned functions by utilizing the hidden representations constructed by
previously learned functions in the new tasks. STINT combines multiple
artificial neural networks in a directed acyclic graph, so as to benefit
from previously learned functions in curricula. Given a set of tasks,
and labeled data for each task, we show that simple greedy algorithms
can generate a curriculum that approximates the optimal curriculum.
I'll then talk about how this notion of curricula occurs naturally in
situations such as robot learning, and will present results testing
these algorithms in a mobile robot domain, showing that learning with a
curriculum can significantly reduce the number of examples required to
learn particular novel tasks.